Introduction

With the emergence of new natural language technologies and the continual maturation of existing technologies, intelligent voice technology has moved from the nascent stage to the mature stage, driving its increasing large-scale commercial use. The most prevalent commercial use thereof concerns voice assistants (VAs) provided with artificial intelligence (AI), e.g., voice-controlled programs embedded in other devices such as Siri or stand-alone devices such as TmallGenie (Ali), Xiaoai Classmate (Xiaomi) or Alexa (Amazon; Malodia et al., 2021; Marikyan et al., 2022; McLean et al., 2021). These AI-powered VAs communicate with users in a human-like manner, mimicking humans as closely as possible, and they are constantly evolving to improve their users’ interaction experiences (McLean and Osei-Frimpong, 2019; Pantano and Pizzi, 2020). In recent research, VAs have been identified as key drivers of service innovation with great potential for growth (Malodia et al., 2022; Juniper Research, 2018). This potential can be seen in terms of VA ownership rates or expected market growth rates. For example, in the U.S., more than 30% of households have a VA (Kinsella, 2020); in Japan, the number of households with VAs is expected to exceed 22 million by 2024 (Francis, 2019). Similarly, in China, this market is expected to reach a staggering $20 billion by 2030 (Deloitte Consulting, 2021).

Given this unprecedented growth, some scholars argue that more research is needed to better understand voice-based AI interaction and to propose related best practices and guidelines (Huang et al., 2021; McLean et al., 2021; Xiao and Kumar, 2021). Recent studies have also suggested that service providers need to understand the benefits of different types of AI and adopt distinct types thereof in disparate scenarios, e.g., mechanical AI for standardization, thinking AI for personalization, or feeling AI for relationalization (Huang and Rust, 2021). However, thinking and feeling AI have yet to be fully realized, resulting in a gap between the capabilities of VAs and human assistants (Huang and Rust, 2021). Repetitive or mechanical responses from VAs during VA-user interactions may frustrate users, who in turn may interact less with these VAs or even leave them completely idle. For example, despite achieving some initial success, Amazon’s Alexa seems to have recently fallen out of favor, with people becoming tired of it and no longer using it (Adrianna, 2022). In China, AI-based voice assistants (such as Alibaba’s Tmall Genie and Xiaomi’s Xiao Ai) started relatively late, and they are currently in a stage of rapid development. However, importantly, they may soon face a similar situation to Alexa. Overall, however, reports in China express positive expectations of voice assistants, anticipating increasing interactions between humans and AI-based VAs (Tan, 2021). In this milieu, the strategic pursuit of competitive advantage for their products via distinctive attributes has emerged as a paramount preoccupation for numerous Chinese Virtual Assistant enterprises. Based on this, this study focuses on whether a simple shift in language helps to improve the user’s interaction experience with a VA under existing AI technology.

Despite the substantial amount of research on AI or other forms of AI products, little attention has been given to the question of how users react to a VA’s language. Language, a direct carrier of VA–human interaction, may have different impacts on users. For example, if someone asks a VA to recommend a song for relaxation, the VA could state, “I have found a song for you. It has the highest number of plays and ratings in the last month,” or “I have found a song for you. It is one of the hottest songs recently.” In the former case, the VA provides a very concrete description of the song. In the latter, the VA uses more abstract wording that generalizes its evaluation to an overall impression of the song. These two language styles may have different impacts on users’ subsequent attitudes and behaviors. Previous consumer research has shown that concrete language is preferred in salesperson–consumer communication because it is effective in increasing consumer satisfaction and purchase intention but abstract language is more effective in word-of-mouth communication (Packard and Berger, 2021; Schellekens et al., 2010). This article therefore systematically discusses the impact of abstract or concrete language in the context of AI–human interaction.

The demand for intelligent VAs is growing at a high rate, and VAs are being increasingly used in all aspects of daily life. The above service contexts can be broadly classified into two categories according to their intrinsic attributes utilitarian-dominant or hedonic-dominant (Liu et al., 2022; Prebensen and Rosengren, 2016). Users’ preferences may differ across these different types of service contexts. Specifically, in utilitarian-dominant contexts, they are more focused on accuracy, responsiveness, and compatibility, while in hedonic-dominant contexts, they are more focused on anthropomorphism and affinity (Cadario et al., 2021; Mishra et al., 2021; Yuan et al., 2022; Zhang et al., 2021). Accordingly, consumers’ or users’ trade-offs between pleasure and utility affect their preference for or resistance to different characteristics of AI (Bhargave et al., 2015; Liu et al., 2022; Longoni and Cian, 2022; Yuan et al., 2022). Hence, this article also explores whether the use of abstract or concrete language by VAs in different contexts (that is, utilitarian-dominant or hedonic-dominant) has a distinct effect on the receivers.

Utilitarian contexts are characterized by an emphasis on function and efficiency, while hedonic contexts emphasize social and emotional needs (Jiang and Lu Wang, 2006; Yuan et al., 2022). Furthermore, concrete language focuses on direct experiences (e.g., precise data or operational methods; Elliott et al., 2015; Shani-Feinstein et al., 2022). Abstract language is often associated with AI anthropomorphism because of its high construal level (e.g., figurative language or high empathic responses; Choi et al., 2019; Lv et al., 2022). This article thus suggests that there is a congruency effect and demonstrates how the match between the language styles and VA service contexts may improve users’ evaluations of them. Specifically, we predict that in utilitarian-dominant service contexts, users react more positively to VAs with concrete language and exhibit a higher continuous usage intention; in contrast, in hedonic-dominant service contexts, users prefer VAs with abstract language. It is also proposed that processing fluency mediates this congruity effect, as the former has been identified as an important result that is increased by the matching of psychological distance (distant vs. close) to information construal level (concrete vs. abstract; Connors et al., 2021).

Overall, this research contributes to the literature through three studies, as there has been limited research on whether language styles and VA service contexts can predict users’ attitudes. Hence, the findings advance the understanding of AI–human interactions within the natural language domain. Furthermore, from a managerial perspective, this research provides actionable insights for managers into how to upgrade the language systems of their VA products to successfully move users through a shift to AI and into optimizing the user experience in a variety of specific service contexts.

Literature review

The utilization of VA

A VA is a software agent that analyses human voice commands and responds with a synthesized voice (Hoy, 2018). VAs are usually in a continuous listening state and can be activated by specific wakeup terms—for example, “Hey, Alexa” or “Hey, Xiaoai” (Xiaomi). Once activated, VAs begin to interact with users in real time, not only for hedonistic purposes (e.g., telling jokes to please a user), but also for utilitarian purposes (e.g., searching for specific information, placing an online order) (Malodia et al., 2022). The development of VAs has greatly changed the mode of human–AI interaction, as users can now perform many tasks through voice commands alone, making their work and life more convenient (Alepis and Patsakis, 2017; Strayer et al., 2017).

Research has shown that the utilitarian benefits, symbolic benefits, and social benefits of VAs can effectively promote user adoption but perceived privacy risks act as a disincentive (McLean and Osei-Frimpong, 2019). Moreover, offering a more personalized value proposition, such as social identity, convenience, perceived playfulness, perceived usefulness, or personification, can effectively engage consumers with VAs (Malodia et al., 2021). In addition, through numerous other specific drivers, VAs can support individuals in performing work-related tasks or serve as new channels for brands to promote consumer brand engagement (Marikyan et al., 2022; Marinova et al., 2017; McLean et al., 2021). On the other hand, the antecedents of consumers’ refusal to use VAs, such as cognitive biases, nudging, consumer inertia, and procrastination, have also been identified (Malodia et al., 2022).

Based on the above, whether consumers/users use VAs can be explained by their technology use or salient factors that reflect the capabilities/value of VAs. However, limited attention has been given to language, the primary channel for interaction between VAs and humans. Nevertheless, previous research on consumer behavior has provided preliminary evidence that language style plays a very important role in interaction with consumers (Packard and Berger, 2021; Schellekens et al., 2010). This article therefore focuses on the potential impact of the VA language style on a more micro level than has been thus far explored.

Language style in AI–human interaction

The language of artificial intelligence is a technical representation designed and implanted through programming that can be modified and enriched in real-time by program developers. Some studies have shown that AIs with higher levels of communication skills (e.g., tailored responses and response varieties) lead to higher perceived social presence and partner engagement (Schuetzler et al., 2020). More relevant to this research is the growing stream of work on the role of language style in service encounters with AI. Warmth and competence are two important aspects thereof (e.g., Lv et al., 2022; Xu et al., 2022).

In the social judgment literature, warmth and competence are two basic dimensions of interpersonal judgment (Fiske et al., 2002, 2007). Warmth refers to sociability, kindness, and dependability (Fiske et al., 2007). Competence, related to characteristics such as capability and skill, refers to an individual’s ability, intelligence, and skillfulness (Kim et al., 2019). Human warmth is a socially desirable quality, whereas competence is a functional and utilitarian quality (Yzerbyt et al., 2008). Although both warmth and competence are crucial, prior researchers have clearly placed greater emphasis on warmth or aspects connected to it (Kull et al., 2021; Kumar et al., 2022; Lv et al., 2022; Tsai et al., 2021; Xu et al., 2022). Thus, brand engagement increases when chatbots open a conversation with a warm (vs. competent) message (Kull et al., 2021). User satisfaction can be increased through a socially oriented communication style or highly empathetic responses, both of which are related to warmth (Kumar et al., 2022; Tsai et al., 2021; Xu et al., 2022). In addition, a socially oriented communication style can mitigate the negative effects of AI–human interaction (Zhou et al., 2022). On the other hand, from the perspective of competence, consumers may respond more favorably to human service agents who use literal (vs. figurative) language (Choi et al., 2019). Clarification AI is also perceived to be competent, as consumers’ opinions thereof are identical to those of error-free AI (Sheehan et al., 2020). Together, these two dimensions of warmth and competence thus encompass nearly all the ways individuals describe others (Fiske et al., 2007). This suggestion is also valid in the realm of AI–human interaction, although few studies have investigated both dimensions simultaneously.

Recent marketing research has examined the combined influence of a message’s construal level (abstract vs. concrete) and related elements (e.g., psychological distance, specific scenarios) on consumers in a more integrated framework via construal-level theory (Connors et al., 2021; De Angelis et al., 2017; Packard and Berger, 2021; Shani-Feinstein et al., 2022). Since concrete and abstract language styles have the respective qualities of competence and warmth, they can be used to indirectly map competence and warmth. Moreover, as the two extremes on the language style continuum, they are more susceptible to manipulation in practice. This article therefore attempts to provide new insights into AI–human interaction by discussing the role of abstract vs. concrete language styles in the AI domain.

Abstract vs. concrete language

This distinction between abstract and concrete language extends from the linguistic category model (LCM; Semin and Fiedler, 1988), a frequently used framework when studying the language that people use to describe interpersonal behavior. That is, the primary distinction between abstract and concrete language is the degree to which abstract verbs and predicates are utilized to express experiences or events. This distinction can be broken down into four categories (Maass et al., 1989; Semin and Fiedler, 1988). For instance, a person who observes A hitting B may state the following: (a) “A hits B”; (b) “A hurts B”; (c) “A hates B”; or (d) “A is aggressive.” Concerning the two extremes, the first is the most concrete level, clearly verifiable and leaving little room for debate; the last is the most abstract level, prone to debate because it generalizes this conduct to the level of A’s personality.

Based on LCM, marketing scholars have already investigated when customers prefer to hear product information expressed in abstract language (vs. concrete language) and when abstract language (vs. concrete language) is more likely to impact its recipients’ attitudes and product purchase intention. In particular, consumers tend to use abstract language when describing product experiences that are consistent with their product attitude and to use concrete language when describing product experiences that are inconsistent with their product attitude. Among receivers, abstract language in positive word-of-mouth leads to better product attitudes and purchase intentions (Schellekens et al., 2010). Furthermore, consumers with a persuasive purpose (e.g., the goal of convincing other consumers to accept a message sender’s point of view) tend to use more abstract (concrete) language to describe their positive (negative) product experiences than consumers without such a purpose (Schellekens et al., 2013). In addition, abstract language is more effective than concrete language among receivers with high prior knowledge (De Angelis et al., 2017). However, abstract language is not always more effective; consumers also favor concrete language in specific situations. Investors, for instance, are significantly more willing to invest in a company when its prospectus emphasizes concrete language rather than abstract language because concrete language increases their comfort in evaluating their target (Elliott et al., 2015). Customers also report higher levels of satisfaction, greater willingness to make purchases, and increased overall spending when employees speak to them concretely (Packard and Berger, 2021). However, some scholars have elaborated on the differences between concrete and abstract language in a more integrated framework. Specifically, a match between informational frames and mental representations can lead to more positive results, such as a match between the “how” (concrete) and “why” (abstract) messages and the temporal frame (Kim et al., 2009); between the “how” (concrete) and “why” (abstract) messages and the speed frame (Shani-Feinstein et al., 2022); and between the construal level of a brand (abstract vs. concrete) and customers’ psychological distance from that brand (close vs. distant, Connors et al., 2021).

Given that LCM provides clear explanations and examples of abstract and concrete language, these types of language are easier to distinguish and manipulate than warm (vs. competence) or socially oriented (vs. task-oriented) language styles (Roy and Naidoo, 2021; Xu et al., 2022). On the other hand, the continuous enrichment of service scenarios requires voice assistants to adopt a more adaptable language strategy to effectively respond to users’ various requests. Extending this stream of research to the present study, we argue that users’ trade-offs in service context attributes (hedonic or utilitarian) may influence their preference for VA language style (abstract or concrete). In the following section, we examine this reasoning.

Congruity between VA language style and service context

The mainstream discussion has already compared humans to virtual assistants (or other forms of AI, such as algorithms or robots). The main view suggests that consumers’ preferences for AI or humans are dynamically changing and that the context of service is an important influencing factor. In their study, Longoni and Cian (2022) find that in the context of hedonic consumption, consumers prefer recommendations from humans but in the context of utilitarian consumption, they prefer AI recommendations. Garvey et al. (2023) show that when receiving bad news, consumers’ reactions are less negative towards AI than humans but that when receiving good news, consumers’ reactions are more positive towards humans than AI. Additionally, factors such as the objectivity of a task, the source of service error (human or algorithm), consumer emotion (anger or non-anger), and service outcome (favorable vs. unfavorable) influence consumers’ preferences for humans or AI (Castelo et al., 2019; Srinivasan and Sarial-Abi, 2021; Crolic et al., 2022; Yalcin et al., 2022). However, regarding VAs, some scholars argue that since artificial intelligence is designed to mimic human interactions and create human‒machine interactions, concreteness during such interactions is crucial. Indeed, concreteness can serve as an important antecedent to perceived usefulness, thereby enhancing the connection between consumers and VAs (Malodia et al., 2022). In the field of communication between businesses and consumers, using concrete (low-level explanatory) language when engaging consumers can effectively increase their satisfaction (Packard and Berger, 2021). However, in certain situations, abstract (high-level explanatory) language can also yield unexpected effects (Schellekens et al., 2010; De Angelis et al., 2017). Building on the above studies, Connors et al. (2021) propose that communication effectiveness between brand and consumer improves when the level of information conveyed by a brand matches the psychological distance of a consumer. Hence, we suggest that in the interaction between humans and VAs, businesses could benefit from utilizing VAs as a medium for dynamically changing language styles based on service context.

Recent studies have also shown that people’s attitudes toward AI are influenced by the difference between hedonistic and utilitarian service situations (Liu et al., 2022; Longoni and Cian, 2022; Wien and Peluso, 2021). Notably, the services provided by companies to consumers in different service contexts often include both hedonic and utilitarian values, while the distinction between these two values is not absolute (Okada, 2005). Rather than a bimodal system, these service contexts might thus be viewed as a continuum from high hedonic to high utilitarian (Parsa et al., 2020). Based on the characteristics of VAs, this research therefore discusses consumers’ preferences for language in two rather common contexts, namely, the hedonic-dominant context and the utilitarian-dominant context. In the utilitarian-dominant context, consumers are more concerned with utilitarian value, often reflecting rationality, nonsensory attributes, instrumentality, functionality, and cognitively driven (Batra and Ahtola, 1991; Hirschman and Holbrook, 1982; Botti and McGill, 2011). In contrast, in the hedonic-dominant context, consumers are more concerned with hedonic value, typically experienced via emotions associated with products and sensory enjoyment and thus characterized by an abundance of emotions and primarily motivated by emotional factors (Batra and Ahtola, 1991; Hirschman and Holbrook, 1982; Botti and McGill, 2011). Specifically, utilitarian-dominant contexts exhibit characteristics such as functional quality and monetary value, whereas hedonic-dominant contexts encompass elements of social interaction, emotional experience, and cognition (Prebensen and Rosengren, 2016). The utilitarian value of AI in AI–human interactions can therefore be enhanced with accuracy, responsiveness, and compatibility, whereas AI’s hedonic value can be fostered through anthropomorphism and affinity (Yuan et al., 2022). Based on these characteristics, researchers have discussed and analyzed specific service contexts for AI. For example, destination recommendations are hedonic-dominant, while insurance advice is utilitarian-dominant (Liu et al., 2022).

Notably, concrete language is rather specific and contextualized, easily visualized, and does not over-explain (e.g., using more numbers; Lundholm et al., 2014). In contrast, abstract language contains more information and often involves more subjective emotional expressions (Schellekens et al., 2010). According to construal-level theory, the greater the psychological distance between an object and a person is, the more likely the object is perceived on a higher level of abstraction. In contrast, psychologically close objects are represented by more concrete, low-level construals (Lee and Labroo, 2004; Trope et al., 2007; Trope and Liberman, 2010). Low-level construals thus tend to be presented in concrete language styles, whereas high-level construals are typically manipulated with abstract language styles. For instance, companies can present messages with different construal levels (abstract vs. concrete) to consumers with disparate psychological distances (distant vs. close) from their brands to obtain better brand evaluations from them (Connors et al., 2021). Furthermore, the consistency effect between the construal level and relevant information frame can also result in a positive evaluation. Indeed a match of gain (loss) frame to an abstract (concrete) mindset can lead to improved conservation behavior (White et al., 2011), while a match between consumer type (problem-focused vs. emotion-focused) and construal level (concrete vs. abstract) renders health messages more persuasive (Han et al., 2016). Moreover, the speed–abstraction effect demonstrates that people’s perception of speed causes them to experience different mental representations during decision-making (Shani-Feinstein et al., 2022). These findings also suggest the possibility of a congruity effect between construal level and specific VA service context, that contributes to improvements in users’ service experiences.

In sum, people use VAs in utilitarian-dominant contexts to solve problems via an emphasis on precision and professionalism. In hedonic-dominant contexts for entertainment or relaxation, a degree of uncertainty may be more captivating to users. Accordingly, this study proposes that in hedonic-dominant service contexts, users prefer VAs with abstract language but that in utilitarian-dominant services, users prefer VAs with concrete language. Thus, we hypothesize the following:

H1a. In hedonic-dominant service contexts, consumers/users exhibit a higher continuous usage intention towards VAs that use abstract language than those that use concrete language.

H1b. In utilitarian-dominant service contexts, consumers/users exhibit a higher continuous usage intention towards VAs that use concrete language than those that use abstract language.

The mediating role of processing fluency

The concept of processing fluency is usually defined as the ease of processing information or meaning; the greater the degree of processing fluency is, the easier it is to process information or evaluate meaning (Alter and Oppenheimer, 2008; Lee and Labroo, 2004). Fluency is thus the result of many congruity effects, based on construal level, e.g., the match between context (promotion-focused vs. prevention-focused) and message frame (gain vs. loss) (Lee and Labroo, 2004); between construal level (concrete vs. abstract) and message frame (gain vs. loss) (White et al., 2011); and between construal level (concrete vs. abstract) and psychological distance (close vs. distant) (Connors et al., 2021). Furthermore, processing fluency can influence message persuasion (Lee and Aaker, 2004), recycling intention and behavior (White et al., 2011), as well as choice (Novemsky et al., 2007), liking (Allard and Griffin, 2017) and brand evaluation (Connors et al., 2021). Therefore, when a specific context is consistent with the construal level, the message may be easier to process. In the context of this research, we therefore expect that matching service context to VA language improves a user’s perception of fluency, which in turn influences his or her evaluation of the voice assistant. Thus, we hypothesize the following:

H2. Processing fluency mediates the congruity effect of VA language style (concrete vs. abstract) and service context (utilitarian-dominant vs. hedonic-dominant) on users’ continuous usage intention.

The overall research model is illustrated in Fig. 1, as follows:

Fig. 1: Research model.
figure 1

The model shows the relationship between the main variables. The arrows indicate the mechanism of influence between the variables.

Full size image

Overview of the studies

In a series of three studies, we employed various scenarios and different samples to test our hypotheses. In the pilot study, we verified whether the VA service contexts can be defined as utilitarian-dominant or hedonic-dominant. In Study 1, we verified the congruity effect of the language style and service context of VAs on users’ evaluation of them. Study 2 repeatedly tested H1 and assessed the mediated effect of processing fluency via new context materials, separately, to verify the stability of the results. The pilot study, study 1 and 2 employed online experiments with participants recruited through Credamo (a professional research data platform from China, similar to Amazon Mechanical Turk) from various provinces in China (see Fig. 2 for distribution of participants). Study 3 was a field experiment that tested our hypotheses using real VAs’ images and voices, with participants mainly drawn from Sichuan Province.

Fig. 2
figure 2

Distribution of participants (pilot study, study 1 and 2; Total number: 967).

Full size image

Pilot study

Based on a survey of relevant VA products and related literature studies, we designed 4 VA service contexts: a music recommendation, a movie recommendation, online shopping, and financial investment. The financial investment context was adapted from Liu et al. (2022). Research has shown that hedonic and utilitarian attributes are important considerations for consumers in their interactions with AI (Longoni and Cian, 2022). The primary value and objective of utilitarian services are to offer convenience, practicality, affordability, and efficiency, with the goal of minimizing monetary waste. Conversely, in hedonic-dominant service contexts, the central value resides in attaining pleasurable sensations, enjoyment, and delightful experiences that evoke feelings of joy, excitement, and personal preference (Botti and McGill, 2011; Ryu et al., 2010). If this is the case, users may prefer to interact with a VA that uses abstract language in a hedonic-dominant context, but they may prefer a VA that uses concrete language in a utilitarian-dominant context. Considering that online shopping and financial investment are relatively important matters in daily life, both functional and utilitarian, and that music and movie recommendations fall under pleasurable home experiences, which are entertaining and hedonistic, online shopping and financial investment were considered utilitarian-dominant service contexts in Study 1 and Study 2, respectively, but music and movie recommendations were regarded as hedonic-dominant service contexts.

Method

The purpose of the pilot study was to verify that our classification of the four specific service contexts stated above was consistent with the actual perceptions of VA users. A total of 240 participants (40.8% male; Mage = 28.61) were recruited via Credamo and randomly assigned to one of the four contexts (i.e., music recommendation, movie recommendation, online shopping, and financial investment; please see Appendix A for details). Each participant received $0.15 as monetary compensation. Their perceptions of these service contexts were measured using three items from Liu et al. (2022) and Voss et al. (2003). One example is, “How do you rate the service in terms of fun and usefulness?” (−3 = More useful than fun, 0 = Equally useful and fun, 3 = More fun than useful; please see Appendix C for details). Each score indicated whether the service context was utilitarian-dominant or hedonic-dominant.

Results

The results showed significant differences in participants’ perceptions of the above four contexts in terms of their utilitarian and hedonic attributes (F(3,236) = 149.75, p < 0.001). A post-hoc analysis demonstrated that the four service contexts could be divided into two subcategories, with the music and movie recommendations being significantly more hedonic than the other two. Specifically, the music recommendation (M = 1.66) was perceived as more hedonic than online shopping (M = − 2.07; p < 0.001), and the movie recommendation (M = 0.47) was regarded as more hedonic than financial investment (M = − 2.45; p < 0.001). Therefore, the music recommendation and online shopping were used as the scenarios for Study 1 but movie recommendation and financial investment were adopted as the scenarios for Study 2.

These results also show that service contexts can be generally classified as utilitarian-dominant or hedonic-dominant during VA–human interaction. For example, a music recommendation is hedonic-dominant, in contrast to online shopping, which is utilitarian-dominant. Based on the findings of this pilot study, Study 1 and Study 2 therefore specifically evaluated the congruity effect of language style and service context.

Study 1

Method

Participants and design. This study examined the interaction effect between the perception of language style (abstract vs. concrete) and service context (hedonic-dominant vs. utilitarian-dominant) on continuous usage intention. Study 1 employed a 2 (perception of VA’s language style: abstract vs. concrete) × 2 (service context: hedonic-dominant vs. utilitarian-dominant) between-subjects design. For this study, 380 people (56.3% male; Mage = 30.24 years) recruited from Credamo participated. Each participant received $0.15 as monetary compensation.

Procedures and materials. For the manipulation of service context, participants in the hedonic-dominant (vs. utilitarian-dominant) service contexts were asked to imagine a scenario in which they were asking the AI to recommend good songs for them to relax (vs. searching shopping sites and recommending related products). A detailed description of the scenarios and the AI voice assistant is provided in Appendix A. Next, they were presented with a set of simulated conversations between the user and the VA. The difference between these two between-subjects conditions was the language style of the VA’s response (abstract vs. concrete; see Appendix B for details). After consideration of the scenarios, their perceptions of the VA’s language style were measured using a single item adapted from Packard and Berger (2021): “How concrete or abstract was the VA’s reply?” (1 = More abstract than concrete, 4 = Equally abstract and concrete, 7 = More concrete than abstract). A higher score indicated that the language style was more concrete than abstract. They were then asked to indicate their perceptions of these service contexts, measured similarly to the pilot study.

Next, we measured the key dependent variables. Continuous usage intention was measured using a three-item, seven-point self-report scale, including the following statements: “I’ll try the VA service again, not just abandon it,” “I’ll continue to use this VA service, not calling for anyone’s help,” and “If possible, I’ll stop using this VA service” (1 = strongly disagree, 7 = strongly agree; Bhattacherjee, 2001). Finally, they answered some basic demographic questions.

Results

Manipulation check

Confirming the effectiveness of the manipulation, a t-test revealed that those in the hedonic condition felt more hedonic than those in the utilitarian condition (Mhedonic = 5.12, SD = 1.89 vs. Mutilitarian = 1.92, SD = 1.22, t (323) = 19.55, p < 0.001). Furthermore, those in the concrete condition felt more concrete about the VA’s response than those in the abstract condition (Mconcrete = 6.25, SD = 0.85 vs. Mabstract = 4.10, SD = 2.06, t(251) = 13.24, p < 0.001).

Continuous usage intention

The results of a 2 (perception of VA’s response: abstract vs. concrete) × 2 (service context: hedonic-dominant vs. utilitarian-dominant) ANOVA yielded a significant interaction effect (F(1, 376) = 26.49, p < 0.001) and a main effect of language style (F(1, 376) = 3.94, p < 0.05). The main effect of service context was not significant (p > 0.50). In general, participants’ continuous usage intention was higher in the concrete response condition than in the abstract response condition (Mconcrete = 5.51 vs. Mabstract = 5.29, F(1,376) = 3.94, p < 0.05). Follow-up simple effects (see Fig. 3) revealed that in utilitarian-dominant service contexts, participants’ continuous usage intention was significantly higher in the concrete condition (Mconcrete = 5.81) than in the abstract condition (Mabstract = 5.01; F(1, 376) = 25.42, p < 0.001). On the other hand, in hedonic-dominant service contexts, participants’ continuous usage intention was significantly higher in the abstract condition (Mabstract = 5.56) than in the concrete condition (Mconcrete = 5.20; F(1, 376) = 5.00, p < 0.05).

Fig. 3: Interaction effect between the perception of VA language style and service context on continuous usage intention.
figure 3

Note: * denotes p < 0.05, *** denotes p < 0.001.

Full size image

Study 2

Study 1 examined the impact of the congruity effect on continuous usage intention. Two of the more common human-VA interaction contexts (i.e., music recommendation and online shopping) were tested in Study 1. Study 2 thus further examined this congruity effect in two more complex scenarios and tested the mediating role of processing fluency. New scenarios and experimental materials were introduced to enhance the generalizability of the study. By comparing preferences for VA language style in a movie recommendation and financial investment, Study 2 attempted to show that the abstract language style is favored in a hedonic-dominant context (i.e., the movie recommendation) but that the concrete language style is favored in a utilitarian-dominant context (i.e., financial investment). In addition, processing fluency should mediate this congruity effect.

Method

Participants and design. This study also employed a 2 (perception of VA’s language style: abstract vs. concrete) × 2 (service context: hedonic-dominant vs. utilitarian-dominant) between-subjects design. A total of 348 people recruited from Credamo participated in this study (50.90% male; Mage = 27.51 years). Each participant also received $0.15 as monetary compensation.

Procedures and materials. Study 2 followed basically the same procedures as Study 1, but it used different service contexts and dialog materials. For the manipulation of service context, participants were randomly assigned to one of two service contexts: a movie recommendation or financial investment. A detailed description of the scenarios and the AI voice assistant is provided in Appendix A. In the different scenarios, participants received responses in a distinct language style from the VA (abstract vs. concrete; see Appendix B for details). After reading the scenario information and dialog material, the participants expressed their continuous usage intention, as in Study 1. Then they were asked to complete a three-item measure of processing fluency on a seven-point scale (e.g., “How easy was it to process the VA’s reply?”; 1 = very difficult, 7 = very easy; Lee and Aaker, 2004). Finally, participants provided their basic demographic information.

Results

Manipulation check

Confirming the effectiveness of the manipulation, a t-test revealed that those in the hedonic condition felt more hedonic than those in the utilitarian condition (Mhedonic = 4.74, SD = 1.94 vs. Mutilitarian = 2.08, SD = 1.42, t(313) = 14.57, p < 0.001). Furthermore, those in the concrete condition felt more concrete about the VA’s response than those in the abstract condition (Mconcrete = 5.85, SD = 1.17 vs. Mabstract = 4.07, SD = 1.86, t(286) = 10.64, p < 0.001).

Processing fluency

The results of a 2 (perception of VA’s response: abstract vs. concrete) × 2 (service context: hedonic-dominant vs. utilitarian-dominant) ANOVA revealed only a significant interaction effect (F(1, 344) = 19.83, p < 0.001). No other main effects were significant (p > 0.50). Further analysis (see Fig. 4) showed that in utilitarian-dominant service contexts, the participants had a higher perception of processing fluency in the concrete language condition (Mconcrete = 5.83) than in the abstract language condition (Mabstract = 5.27; F(1, 344) = 11.19, p < 0.01) but that in hedonic-dominant service context, the participants had a higher perception of processing fluency in the abstract language condition (Mabstract = 5.78) than in the concrete language condition (Mconcrete = 5.27; F(1, 344) = 8.73, p < 0.01).

Fig. 4: Interaction effect between the perception of VA language style and service context on fluency.
figure 4

Note: ** denotes p < 0.01.

Full size image

Continuous usage intention

The results of a 2 (perception of VA’s response: abstract vs. concrete) × 2 (service context: hedonic-dominant vs. utilitarian-dominant) ANOVA revealed a significant interaction effect (F(1, 344) = 20.97, p < 0.001) and a main effect of service context (F(1, 344) = 5.11, p < 0.05). The main effect of language style was not significant (p > 0.50). In general, continuous usage intention was significantly higher in the hedonic context than in the utilitarian context (Mhedonic = 5.54 vs. Mutilitarian = 5.28, F(1344) = 5.11, p < 0.05). Further analysis (see Fig. 5) showed that in utilitarian-dominant service contexts, the participants had a higher continuous usage intention towards the VAs with a concrete language style (Mconcrete = 5.53) than towards those with an abstract language style (Mabstract = 5.03, F(1, 344) = 9.68, p < 0.001) but that in hedonic-dominant service contexts, the participants had a higher continuous usage intention towards the VAs with an abstract language style (Mabstract = 5.81) than towards those with a concrete language style (Mconcrete = 5.26; F(1, 344) = 11.31, p < 0.01).

Fig. 5: Interaction effect between the perception of VA language style and service context on continuous usage intention.
figure 5

Note: ** denotes < 0.01, *** denotes < 0.001.

Full size image

Moderated mediation

To assess our moderated mediation hypotheses, we used the PROCESS macro (Hayes, 2018), which is often used to test indirect effects. PROCESS uses bootstrapping, a nonparametric, computationally intensive technique to repeatedly sample from the data set and make statistical inferences from each sample. If the 95% confidence interval (CI) values do not contain zero, we can conclude that the inference is statistically significant (Preacher and Hayes, 2008; Shrout and Bolger, 2002). We used PROCESS Model 8 with 5000 bootstrap iterations to test the significance of the unstandardized indirect effects described in Hypothesis 2. Concrete and abstract language were coded as 0 and 1, respectively; utilitarian-dominant and hedonic-dominant scenarios were also coded as 0 and 1, respectively. These results suggested that fluency mediates the focal relationship since the indirect effect of the highest-order interaction (language style × service context) through fluency was significant (index = 0.71, SE = 0.19; 95% CI = [0.3688, 1.1008]). That is, the effect of language style on continuous usage intention through processing fluency is conditional on service context. In hedonic-dominant service contexts, the results showed a significant indirect effect (β = 0.34, 95% CI = [0.1164, 0.5855]) but in utilitarian-dominant contexts, the direction of the indirect effect was reversed (β = −0.37, 95% CI = [ − 0.6357, −0.1401]).

Study 3

Studies 1 and 2 demonstrated a congruity effect during human-VA interaction, resulting in a higher continuous usage intention towards VAs. Perceived fluency thus plays a mediating role in this relationship. Furthermore, this conclusion is valid across different contexts, providing some evidence for its generalizability.

Nevertheless, whether this relationship is truly causal remains undetermined. In the above studies, we did not employ actual VAs but instead relied on textual descriptions. To address this limitation, study 3 was an on-site investigation of real-time human responses to VAs, further testing the underlying mechanisms and discussing some potential alternative explanations.

Method

Participants and design. This study also employed a 2 (perception of VA’s language style: abstract vs. concrete) × 2 (service context: hedonic-dominant vs. utilitarian-dominant) between-subjects design. A total of 165 first and second-year students from a specialized university in Chengdu participated in this survey. Four individuals were excluded due to failing the attention check, resulting in an actual participation count of 161 people (42.9% male; Mage = 19.37 years). Each participant received 3 yuan as monetary compensation.

Procedures and materials

Study 3 employed the same context materials as Study 1, namely music recommendations and online shopping. A detailed description of the scenarios and the AI voice assistant is provided in Appendix A. Simultaneously, we utilized relevant AI tools to design the appearance and voice of the VA and synthesized them into videos. To mitigate potential anthropomorphism effects on human responses, the VA’s appearance and voice were intentionally infused with more typically robotic characteristics (see Appendix D for VA appearance). Participants were randomly assigned to different scenarios and gave verbal commands to the VA on their computer screen based on material prompts. Subsequently, the investigators played the VA’s response video for the participants. In these different scenarios, participants received responses in a distinct language style from the VA (abstract vs. concrete; see Appendix B for details). Afterwards, the participants expressed their continuous usage intention, and processing fluency, as in Study 2. Finally, we adapted 9 additional items from prior research to address alternative explanations based on the accuracy or usefulness of VA responses (see Appendix C for measurement items; Yuan et al., 2022; Davis, 1989).

Results

Manipulation check

Confirming the effectiveness of the manipulation, a t-test revealed that those in the hedonic condition felt more hedonic than those in the utilitarian condition (Mhedonic = 5.71, SD = 1.62 vs. Mutilitarian = 2.33, SD = 1.47, t(159) = 13.76, p < 0.001). Furthermore, those in the concrete condition felt more concrete regarding the VA’s response than those in the abstract condition (Mconcrete = 5.28, SD = 1.68 vs. Mabstract = 3.88, SD = 2.06, t(152) = 4.74, p < 0.001).

Processing fluency

The results of a 2 (perception of VA’s response: abstract vs. concrete) × 2 (service context: hedonic-dominant vs. utilitarian-dominant) ANOVA yielded a significant interaction effect (F(1, 157) = 43.45, p < 0.001) and a main effect of service context (F(1, 157) = 21.98, p < 0.05). The main effect of language style was not significant (p > 0.50). Further analysis (see Fig. 6) showed that, in utilitarian-dominant service contexts, the participants had a higher perception of processing fluency in the concrete language condition (Mconcrete = 5.20) than in the abstract language condition (Mabstract = 4.25; F(1, 157) = 17.12, p < 0.001) but that in hedonic-dominant service contexts, the participants had a higher perception of processing fluency in the abstract language condition (Mabstract = 6.04) than in the concrete language condition (Mconcrete = 4.90; F(1, 157) = 27.23, p < 0.001).

Fig. 6: Interaction effect between the perception of VA language style and service context on fluency.
figure 6

Note: *** denotes p < 0.001.

Full size image

Continuous usage intention

The results of a 2 (perception of VA’s response: abstract vs. concrete) × 2 (service context: hedonic-dominant vs. utilitarian-dominant) ANOVA revealed a significant interaction effect (F(1, 157) = 75.27, p < 0.001) and a main effect of service context (F(1, 157) = 26.81, p < 0.001). The main effect of language style was not significant (p > 0.50). Further analysis (see Fig. 7) showed that in utilitarian-dominant service contexts, the participants had a higher continuous usage intention towards the VAs with a concrete language style (Mconcrete = 5.04) than towards those with an abstract language style (Mabstract = 3.95, F(1, 157) = 31.51, p < 0.001) but that in hedonic-dominant service contexts, the participants had a higher continuous usage intention towards the VAs with an abstract language style (Mabstract = 5.80) than towards those with a concrete language style (Mconcrete = 4.57; F(1, 157) = 44.77, p < 0.01).

Fig. 7: Interaction effect between the perception of VA language style and service context on continuous usage intention.
figure 7

Note: ** denotes p < 0.01, *** denotes p < 0.001.

Full size image

Moderated mediation

Study 3 also included a conditional process analysis with the PROCESS macro (Model 8; Hayes, 2018), using a bootstrap procedure (5000 draws) to construct bias-corrected confidence intervals. Concrete and abstract language were coded as 0 and 1, respectively; utilitarian-dominant and hedonic-dominant scenarios were also coded as 0 and 1, respectively. These results suggested that fluency mediates the focal relationship since the indirect effect of the highest-order interaction (language style × service context) through fluency was significant (index = 0.76, SE = 0.21; 95% CI = [0.37, 1.24]). That is, the effect of language style on continuous usage intention through processing fluency is conditional on service context. In the hedonic-dominant service context, the results showed a significant indirect effect (β = 0.42, 95% CI = [0.20, 0.67]) but in utilitarian-dominant contexts, the direction of the indirect effect was reversed (β = −0.34, 95% CI = [ − 0.62, −0.12]; see Table 1 for details).

Table 1 Moderated mediation model (N = 161).
Full size table

Alternative explanations

Analyses confirmed no difference between conditions for perceived response accuracy (Mconcrete = 5.13, SD = 1.15 vs. Mabstract = 5.10, SD = 1.18, t(159) = 0.19, p = 0.48). Furthermore, to test the explanatory power of perceived response accuracy and usefulness, we examined whether any of these mediate the effect of language and context on continuous usage intention (PROCESS model 8; Hayes, 2018). We removed processing fluency from the model and ran each potential mediator independently. None of the alternatives mediated the effect of language and context on continuous usage intention (accuracy indirect effect = 0.21, SE = 0.14, 95% CI = –0.02, 0.53; usefulness indirect effect = 0.06, SE = 0.12, 95% CI = –0.11, 0.34). Thus, these alternative explanations cannot account for the above results.

Conclusions and general discussion

General discussion

AI voice assistants were originally designed to satisfy users’ hedonic purposes, but as the technology has evolved, they have become increasingly used for utilitarian purposes. Recent studies have demonstrated that applicable scenarios for AI voice assistants frequently form a hedonic-utilitarian continuum rather than a simple binary structure. In addition, the voice assistants on the market tend to use only one language style, but consumers have distinct preferences for certain language styles in different situations. Hence, we propose that it is important to address whether a dynamically changing language strategy can effectively improve users’ evaluations of a voice assistant. We address this issue across three studies and show a mindset-congruency effect whereby hedonic and utilitarian trade-offs determine the preference for the language style (concrete or abstract) of AI voice assistants. The pilot study revealed that the applicability scenarios of AI voice assistants can also be deemed utilitarian-dominant or hedonic-dominant, consistent with the study by Liu et al. (2022). For example, in VA–human interactions, online shopping, and financial investment are utilitarian-dominant contexts, while music recommendations and movie recommendations are hedonic-dominant contexts. Study 1 thus showed that concrete (abstract) language can lead to a higher evaluation of a VA in a utilitarian-dominant (hedonic-dominant) context. Study 2 and 3 further confirmed this congruity effect; that is, users preferred the use of abstract language in a hedonic-dominant context (e.g., movie recommendations) and were more willing to accept concrete language in a utilitarian-dominant context (e.g., financial investment). Additionally, the interaction effect of a VA’s language style and service context on VA evaluation is mediated by processing fluency. Specifically, when language style matches the service context, the perceptions of processing fluency increase, which in turn improves users’ evaluation of the VA.

Theoretical implications

This article makes several theoretical contributions. First, we extend the research on AI–human interactions by addressing the question of whether hedonic/utilitarian trade-offs affect a user’s preference for language style, which in turn improves his or her evaluation of the VA. Limited research has demonstrated that consumers favor AI recommenders in the utilitarian realm but prefer human recommenders in the hedonic realm (Longoni and Cian, 2022). In addition, customer/tourist preferences for robot appearance (warm vs. competent) are also different in these two contexts (Liu et al., 2022). This article therefore expands the research field for AI services by complementing the abovementioned studies.

Second, this research extends previous studies on AI–human interactions, which tend to focus on how people react to a specific language style, by introducing two language styles with distinct construal levels (i.e., concrete and abstract) and by proposing a new language strategy for VAs to improve users’ evaluations. Research on AI–human interactions has already explored the impact of AI language styles from different perspectives (e.g., social-oriented or task-oriented; Choi et al., 2019; Kumar et al., 2022; Lv et al., 2022; Roy and Naidoo, 2021; Xu et al., 2022). The present paper, based on the language effect of AI and research on AI service contexts, has identified a novel congruity effect in AI–human interactions. Whereas our predecessors adopted a distinctive language style to improve people’s evaluations of AI, our findings enrich AI–human interactions by confirming that it is feasible to improve this evaluation by matching the language styles and service contexts of AI voice assistants.

Third, in this article, the characteristics of AI–human interactions are used to deconstruct the internal mechanism of the congruity effect of language style and service context. This reveals the importance of improving people’s processing fluency in AI–human interactions. Previous research has shown that, when people are confronted with different AI language styles, their perceptions of anthropomorphism or their psychological distance from AI play an important role (Kull et al., 2021; Lv et al., 2022; Roy and Naidoo, 2021; Sheehan et al., 2020). Drawing on construal-level theory, this paper explains the mechanism underlying this congruity effect and provides theoretical support for research in the field of AI–human interactions.

Managerial implications

Amid the rapid development of artificial intelligence, the intelligent voice industry has also entered a rapid development stage, whereby the demand for intelligent voice applications in various fields is expanding, leading to smart homes, smart cars, and smart medical services. This trend requires developers in internet companies and intelligent voice technology companies to continuously improve and refine intelligent voice systems to optimize the customer experience. Our insightful findings will help developers or product managers achieve this goal.

The first practical implication of our findings is that developers and product managers should pay attention to the various VA service contexts at the outset of language programming. AI technology allows people to use VAs in an increasing number of scenarios, such as searching for information, shopping, and entertainment. This research indicates that users’ preferences for the construal levels of VAs’ responses vary with the attributes of service contexts. When they feel they are in a utilitarian-dominant service context, users prefer the responses of VAs to be represented more concretely. Therefore, it is important for developers or product managers to be aware of these effects and, where possible, to collect data on the various contexts in which their users interact with VAs and to assess the service attributes of relevant contexts (hedonic-dominant or utilitarian-dominant). On the one hand, such information can be gathered and assessed using big data; on the other hand, it can be collected through online user surveys. Although the former method, is simpler and faster, personalization tends to be ignored; the latter method can thus solve the problem of personalization but has a higher cost. The evaluation of such attributes of various service contexts can enable the dynamic switching of the language styles of VAs. Obviously, there may be scenarios in which neither utilitarian nor hedonic traits can predominate, and it is always appropriate to select the concrete language style in such scenarios (Elliott et al., 2015).

The second managerial implication is that in developing VAs, the development of different types of language packages bears great practical importance. In this research, several pairs of abstract and concrete uses of language in different scenarios, related but easily distinguishable from each other, were designed according to the construal level. Developers can use this work to design language packages with abstract and concrete properties for various VA usage scenarios. That is, concrete languages emphasize the use of numbers and verbs, whereas abstract languages focus more on words with emotional value. In addition, considering the role of fluency in this consistency effect, developers should avoid words or phrases that affect a user’s understanding, such as advanced vocabulary or unfamiliar words; if specialized vocabulary is involved, it is best to provide further explanations.

Limitations and future directions

In this research, both written instructions (Tassiello et al., 2021) and video presentations were employed to simulate interactions between users and VAs in different scenarios, enhancing the generalizability of the results. Notably, the studies employed nonphysical VAs instead of physical VAs, which could have weakened participants’ sense of immersion and perceived authenticity, potentially impacting the validity of the results. Since the literature indicates no significant differences in the effects on certain consumer behaviors (such as pro-environmental behavior intervention) of virtual assistants or physical robots (Tussyadiah and Miller, 2019), the findings in this article may be extended to physical VAs. However, the most ideal situation is using physical VAs to evaluate consumers’ attitudes and responses in real interactions (Tassiello et al., 2021). Real settings in which participants listen to the responses of physical VAs and engage in actual interactions can be employed in future research to further validate the congruity effect proposed in this article.

Moreover, we have focused on the relative degrees of effectiveness of abstract vs. concrete language in this research. However, there are many other language styles that are worth exploring in the context of service encounters, such as affiliative vs. aggressive humor (Béal and Grégoire, 2022). In addition, it would be interesting to explore the effects of different voice types, such as male vs. female or celebrity vs. classic robot. Furthermore, in this study, the theoretical model was only validated in the Chinese context. Future research can thus consider the robustness of the model in a broader range of language contexts, such as English, French, and others. Testing these effects could deepen the understanding of AI–human interactions.

In addition, although we tested the congruity effect of construal level and service context across multiple domains, this effect could be stronger or weaker in certain conditions. For instance, this congruity effect may be moderated by social status, since those in the upper classes are more efficiency-oriented than those in the lower classes, who prefer fun (Kim, 2022). Therefore, the effect of concrete language on people of a low social class may be diminished in a utilitarian-dominant service context. In contrast, the effect of abstract language on people of a high social class may be diminished in a hedonic-dominant service context. Finally, recent studies have indicated that consumers react differently to information from AI than that provided by humans (Garvey et al., 2023). Accordingly, we can shift the focus of research away from VAs to explore consumers’ preferences for AI-based VAs vs. humans in more complex scenarios. Future research could thus more systematically investigate which dimensions of different variables strengthen or weaken this congruity effect.